Frequently asked questions.

Your rating shows how well you play your cards, not whether you win or lose. After each deal we replay the same deal with the AI in your seat, the same cards and the same fellow players. The AI's score shows what was achievable with your cards, and your result is compared to it — so your decision counts, not your luck with the cards.

You start at 5000; higher means you consistently get more out of your cards than expected.

Losing to the AI is normal and doesn't cost you rating by itself — everyone is measured against the same strong yardstick, and you're compared to your own expected level. Winning a round doesn't automatically raise your rating either: you can win a lucky deal without playing optimally, or lose a hard one while playing very well.

The rating uses the neural brain as its yardstick — even when you yourself play against the classic "Tree" brain. See also "What's the difference between the Boom and Neural AI?" and "What does the neural brain have to do with my rating?"

Via the details screen at the end of a round you can see in full detail what the computer would have played.

That's correct. Once you have bought Kraken (or the Trainer), you can activate it yourself in the app settings. Open the settings and scroll down a bit to find the Kraken section.

My wife was used to playing with the "small card with jack" convention: an agreement between two partners. It attaches a conventional meaning to the lead of the first trick if that card is not a trump. If you lead a non-trump low card (7, 8 or 9), this tells your partner:

Partner, I have the Jack, try to get the lead as soon as possible and then lead, preferably, a meld-sensitive trump (Queen, 10) for my Jack.

The advantage: with a trump holding of Jack + King you can still catch the nel (ten) if it sits to the right of the player and his partner can lead the 10. You use the rule that overtrumping is mandatory and avoid the risk that a lead with the Jack only yields the 7, 8 and the Queen.

The convention also has risks: if the partner cannot get the lead because he has no side cards, the opposing side can start trumping in tricks they were not really "entitled to". The other risk: if you do not have the Jack and do not want to draw trumps immediately, you must lead a face card in a non-trump suit, creating a "dangerous hand" situation.

In a future version it will be possible to disable this convention.

The AI only does this when it has to lead while its partner is playing the contract. It assumes its partner must have the jack and will lead meld-sensitively.

Hard to say. What is easier to answer is the signals the AI does respond to:

• Signalling down with a single card
• Signalling down with two cards (the AI makes a random choice here if it has that option)
• Signalling up (e.g. throwing a face card on a partner's trick)
• Throwing an ace on its partner's trick implies it has the 10 of that suit
• If neither the bot nor its partner is playing the contract, and the bot has to lead, it will lead into a singleton or a suit of three or more. The intention is always that its partner takes the trick and returns that suit (to "steal" tricks)
• Small card with jack (see also above)
• If the AI is playing the contract and leads an Ace on the first trick, it has the nel but not the trump jack
• If the AI leads a 10 while the Ace has not yet been played, it most likely also has the Ace

Counter-suit signalling the AI does not do, nor will it recognise it.

A counter-pit is simply NAT (lost) and in this klaverjas app you get no extra points for it. See also article 17 of the KBN game rules:

Article 17 — Only the playing side can qualify for a "march", "pit", "pan" (162 points + 100 meld) for taking all eight tricks.

In some regions you do, but in this app you get only 100 meld for four of a kind (and an honourable mention in your achievements).

No idea. It works perfectly fine. The AI, however, is programmed to keep meld chances for tricks that will (certainly or probably) go to the opponent as small as possible.

That works fine too. Maybe you have very bad cards, just bad luck, or you can't play klaverjas. It sometimes takes even experienced klaverjas players time to play a PIT with this app, but it is certainly possible.

Absolutely not. The app does not cheat. Every decision the AI makes is entirely based on the previous course of play, its partner's signals, the signals the opponents gave each other and the cards still in hand and in play.

There is no peeking and the shuffling of the cards is also completely random:

// shuffle implementation in the app
for (id anObject in kaartSpel) {
    NSUInteger randomPos = arc4random() % ([tmpArray count] + 1);
    [tmpArray insertObject:anObject atIndex:randomPos];
}

Realise too that I have no interest whatsoever in letting the app win against you, or in deliberately making your partner dumber. I prefer satisfied customers who can play a good game of klaverjas over frustrated customers because I deliberately fooled them.

The only difference between US and THEM — that is really YOU.

The beauty of object-oriented programming is that you only have to program an object once and can then instantiate it multiple times. So too in this app: the three computer-controlled players are all three instances of exactly the same object. That means your partner is exactly as smart or dumb as your opponents.

That the opponents seem smarter has to do with the fact that they know exactly how and what is signalled, and that they make no mistakes counting cards. Unintentional signalling to the AI can give strange results: the AI does not account for foolish human players — a signal will therefore always be followed, even if a human would immediately see that it is not wise.

In principle the player with the highest individual meld in hand is awarded the meld; any "stuk" in hand does not count in this comparison. With equal meld, the sequence is considered (Q-K-A then beats 7-8-9), and if those are also equal the card suit is considered.

Suit order (high to low):

Spades · Hearts · Clubs · Diamonds

Of course I am open to nice suggestions to improve the app. Many of the suggestions I have received have by now found their way into the app (new backgrounds, achievements, sounds, adjustable time after a trick, etc.). Got a good idea? Let me know — support@knippie.com.

In Settings → advanced settings → AI brain you pick your opponent. Boom (Tree) is the classic brain that follows fixed rules we wrote ourselves — always built in and ready. Neural is a learning brain that taught itself to play by playing a great many games; on average it plays stronger. Both are tough opponents. See also the "AI opponent" chapter in the manual.

Once. The neural brain is a set of trained model files. Set AI brain to Neural and the app fetches them (small, one-time); after that it works offline. If they're missing, the app offers to re-download them. If an improved brain becomes available later, the app lets you know and you can fetch the new models.

Absolutely. Neural is stronger than Tree on average, but a strong human player will still beat it. It's challenging, not unbeatable. In fact, your well-played games can help us make the brain even better — see "Help improve the AI".

At Monkey and Beginner the opponents deliberately play weaker so you can practise. At Pro the AI plays at full strength. Because measuring on the easy levels would unfairly inflate your rating, your rating only changes at Pro. (Your partner, North, always plays at full strength.)

Those are table-talk lines from the virtual players — short reactions to the game, purely for atmosphere. They're independent of your rating and can be turned off in settings ("Opponent chatter").

Some chat replies are AI-generated and can occasionally be off. See the privacy policy for details.

A voluntary option that helps us train the neural brain. The brain learns most from well-played games by strong players. Switch it on (in Settings → advanced settings) and we may collect your well-played games anonymously — no username or personal details, only how the game went (cards, moves, score). It's off by default; you can turn it off again any time. (Details in the privacy policy.)

Yes. When someone shares a game, you play the exact same cards against the same AI as the sharer, so the comparison is apples-to-apples. If the sharer played the neural brain, so do you. (If the AI version differs, a small note may say the comparison is "indicative".)

A neural brain (a "neural network") is a kind of computer program that learns from examples instead of following fixed rules. Think of learning a language: the classic brain (Tree) memorised a grammar book, while the neural brain picked the language up by listening and playing a great deal. That's why it often has a better feel for the smartest move in a given spot.

By playing an enormous number of games — against itself and against earlier versions — and remembering what worked each time. That way it got better step by step, without us having to spell out every good move. This is called machine learning. (And your well-played games can help with this — see question 19.)

No. It sees only what any player at the table sees: its own cards, the cards played, and the signals. It makes a strong guess based on patterns — it doesn't peek at your hand. (See also "I get the impression the app cheats".)

Tiny differences in how the calculation runs on different devices can make a single move come out slightly differently. It doesn't change how strong it is — it stays a tough opponent.

Your rating uses the neural brain as its yardstick: after your game the strong neural AI replays the same deal to show what was achievable. That's why the rating works even when you yourself play against Tree. See also "How does the rating work?".

In an online table, empty seats can be filled by computer players. Each computer player is run by one of the human players' devices and uses that player's chosen AI brain — usually the classic Tree brain (Neural only if that player has switched it on and downloaded the models). The rating does not apply to multiplayer.

In the Daily Challenge everyone plays against the exact same AI — the classic Tree brain — with the same cards, so the worldwide leaderboard is a fair comparison. The neural brain is deliberately not used here: it needs a download and can come out very slightly differently per device, which would make a global contest unfair.